“We started with all Frontier models. They're great. They're incredibly capable. We use all of them. But when you get into a domain and a task, the big models are good at everything, but they can be beaten by models that maybe don't have the same world knowledge, but have enough to do my job really well. We watched this play out three times. Like when I brought up Nemotron Super, which was what we used for planning, it legitimately beat all the other Frontier models by two points on our evaluation, or we wouldn't have put it in production. And that's a 100 billion parameter model. So it's not tiny, but it's not huge.”
Ask Noah Show · Hands-On Engineering Podcasts · October 2026
Context
Asked whether open source models can beat proprietary ones, the Red Hat CEO answers from Red Hat's own internal agent project, where a mid-sized open model won on a specific task.
Transcript
…of the world because you can shape them. And then that for me, that combination was the, okay, I think I'm unstoppable at this point. And I think there's a spark of that with that's AI to them. Powerful technology. I can shape it. I can put them into agents. I can solve any problem, I would.
Could you make the case that open source genuinely beats the proprietary models, or could you address where maybe there's some opportunity for growth for them?
I think there are two. In my case that I shared, we started with all Frontier models. They're great. They're incredibly capable. We use all of them. But when you get into a domain and a task, the big models are good at everything, but they can be beaten by models that maybe don't have the same world knowledge, but have enough to do my job really well. We watched this play out three times. Like when I brought up Nemotron Super, which was what we used for planning, it legitimately beat all the other Frontier models by two points on our evaluation, or we wouldn't have put it in production. And that's a 100 billion parameter model. So it's not tiny, but it's not huge. So in those cases, and planning for us is it's the me agent that you can ask any business question you can think of, and it knows how to talk to all the organizational agents, specialized questions. It's a complex task, and it beat Frontier models at that. Then if you take the long arc, I think Karpathy, I think it was, had brought this up, not credited to me, but was if you can distill just…
highly biased participant on this because that's also in my budgetary approval. But for me, I use, let's call them like the 20 billion parameter models. To do things that I think most people would say, like that lives in the world of Opus on it. And it does take a little more work and effort, but…
of them, but I see the same power down this path. There's so much you can do, and models just get better every year.
Does AI, or does Red Hat have a story for the person who wants to take control of their entire AI stack, not just reduce some cloud costs?
More from Ask Noah Show
